| Abstract [eng] |
Prostate cancer is one of the most common malignant diseases worldwide. An important component of its diagnosis is the analysis of multiparametric magnetic resonance imaging scans. Although deep learning methods have significantly improved prostate cancer detection and segmentation, this task remains challenging. This dissertation investigates ensemble methods for prostate cancer segmentation with the aim of improving segmentation accuracy and reliability. The study also examines prostate zonal segmentation, as previous research has shown that zonal information enhances the accuracy of cancer detection and segmentation. Two novel deep learning–based pipelines for prostate zonal segmentation are proposed. The most accurate results were achieved using a pipeline based on separate segmentation models for the transition and peripheral zones. Several U-Net architectures incorporating attention mechanisms were also evaluated; however, no statistically significant differences were found among them. Furthermore, the dissertation proposes four novel ensemble methods for prostate cancer segmentation that utilize prostate zonal masks. The best performance was achieved by combining a deep neural network model for peripheral zone cancer segmentation with a model that segments cancerous regions throughout the entire prostate and uses zonal masks as an additional input. Finally, a novel evaluation metric, the tolerant Dice similarity coefficient, is introduced, accounting for acceptable variation in tumor delineations. |
